Kagi LLM Benchmark

Kagi's unpublished reasoning, coding and instruction questions.

Published by

Kagi LLM Benchmark

Category

Reasoning

Index weight

0.5

Models

135

Data as of

9 Sept 2026

Kagi (kagi.com).

1. 1Claude Fable 591.4%
2. 2Claude Fable 588.8%
3. 3GPT-5.588.8%
4. 4Claude Opus 4.888.8%
5. 5Claude Opus 4.683.6%
6. 6Grok 4.583.5%
7. 7Claude Opus 4.780.7%
8. 8Claude Opus 4.580.2%
9. 9Gemini 3 Pro80.1%
10. 10Kimi K2.578.5%
11. 11GPT-5 Pro76.8%
12. 12GLM-575.0%
13. 13Grok 4.2075.0%
14. 14GPT-4o75.0%
15. 15Claude Opus 474.3%

SearchAll providersAI21 LabsAlibabaAmazonAnthropicBaiduByteDanceCohereDeepSeekGoogleInceptionMeituanMetaMicrosoftMiniMaxMistral AIMoonshot AINous ResearchNVIDIAOpenAIStepFunUnknownxAIZhipu AIOpen weights135 of 135

#	Model	Score ↓	Run ↕	Index ↕	Published ↕
1	Claude Fable 5Anthropic	91.4%	–	–	–
2	Claude Fable 5Anthropic	88.8%	–	70.4	–
3	GPT-5.5OpenAI	88.8%	–	66.6	–
4	Claude Opus 4.8Anthropic	88.8%	–	–	–
5	Claude Opus 4.6Anthropic	83.6%	–	62.3	–
6	Grok 4.5xAI	83.5%	–	63.4	–
7	Claude Opus 4.7Anthropic	80.7%	–	–	–
8	Claude Opus 4.5Anthropic	80.2%	–	60.6	–
9	Gemini 3 ProGoogle	80.1%	–	61.6	–
10	Kimi K2.5Moonshot AIopen	78.5%	–	55.3	–
11	GPT-5 ProOpenAI	76.8%	–	60.7	–
12	GLM-5Zhipu AIopen	75.0%	–	57.0	–
13	Grok 4.20xAI	75.0%	–	54.3	–
14	GPT-4oOpenAI	75.0%	–	42.3	–
15	Claude Opus 4Anthropic	74.3%	–	54.4	–
16	Qwen3 5 397BAlibaba	73.7%	–	–	–
17	Grok 4xAI	73.6%	grok-4	57.2	–
18	Claude Opus 4.7Anthropic	73.3%	–	65.5	–
19	GPT-5.2OpenAI	73.3%	–	61.5	–
20	Claude Sonnet 4Anthropic	73.0%	–	54.0	–
21	GPT-5OpenAI	72.7%	gpt-5	59.5	–
22	Qwen3 MaxAlibaba	72.5%	–	52.7	–
23	Claude Opus 4.6Anthropic	72.4%	–	63.1	–
24	o3-proOpenAI	72.1%	–	54.1	–
25	Claude Opus 4.5Anthropic	70.7%	–	58.0	–
26	GPT-5-CodexOpenAI	70.3%	–	58.1	–
27	GPT-5 miniOpenAI	70.3%	–	55.8	–
28	Gemini 2.5 ProGoogle	70.3%	–	54.1	–
29	Qwen3 235B A22BAlibabaopen	69.4%	–	51.7	–
30	DeepSeek R1DeepSeekopen	69.4%	deepseek-r1	48.9	–
31	Claude Haiku 4Anthropic	69.0%	–	–	–
32	o3OpenAI	67.6%	o3	60.3	–
33	o4-miniOpenAI	67.6%	–	54.9	–
34	Gemini 3.1 Flash LiteGoogle	67.2%	–	54.3	–
35	GPT-5.6 SolOpenAI	67.0%	–	65.3	–
36	Qwen3-Next 80B A3BAlibabaopen	66.7%	–	51.0	–
37	URing 1tUnknown	66.4%	inclusionai/ring-1t	48.0	–
38	Grok 4 FastxAI	66.1%	–	53.0	–
39	Kimi K2Moonshot AIopen	64.4%	–	–	–
40	GPT-5.4OpenAI	63.8%	–	63.4	–
41	Kimi K2.5Moonshot AIopen	63.8%	–	56.1	–
42	Gemma-4-31B-ITNVIDIAopen	63.5%	–	55.4	–
43	Qwen PlusAlibaba	63.3%	–	48.4	–
44	GLM-5.2Zhipu AIopen	62.6%	–	57.4	–
45	Step 3StepFun	62.3%	stepfun-ai/step3	51.9	–
46	GPT-5 nanoOpenAI	62.2%	–	51.1	–
47	Grok 3 minixAI	61.3%	–	52.0	–
48	Grok 3xAI	61.3%	grok-3	50.0	–
49	UCogito v2 Llama 405BUnknown	61.1%	–	–	–
50	GLM-5.2Zhipu AIopen	60.0%	–	–	–
51	GLM-4.5VZhipu AIopen	59.8%	–	42.9	–
52	Claude Opus 4Anthropic	59.6%	–	50.7	–
53	gpt-oss-120bOpenAIopen	58.6%	–	47.6	–
54	UCogito v2 Deepseek 671BUnknown	58.4%	–	–	–
55	Claude Sonnet 4.5Anthropic	57.9%	–	52.3	–
56	GLM-4.5Zhipu AIopen	57.9%	–	50.2	–
57	MiniMax-M2MiniMaxopen	57.8%	minimax/minimax-m2	55.1	–
58	ULing 1tUnknown	57.8%	inclusionai/ling-1t	45.1	–
59	DeepSeek V3.1DeepSeekopen	57.4%	–	51.0	–
60	Gemini 2.5 FlashGoogle	56.8%	–	48.1	–

Show 75 more